-
Notifications
You must be signed in to change notification settings - Fork 187
/
c01ffedd-443f-421b-b7ca-492e8187557b.txt
2162 lines (2089 loc) · 107 KB
/
c01ffedd-443f-421b-b7ca-492e8187557b.txt
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
import os
import sys
with open(sys.argv[0]) as f:
code = f.read() # read the code of this file ASAP, for logging
import uuid
import time
import contextlib
from dataclasses import dataclass
from pathlib import Path
import torch
from torch import nn
import torch.nn.functional as F
import torch.distributed as dist
import torch._inductor.config as config
from torch.nn.parallel import DistributedDataParallel as DDP
from torch.nn.attention.flex_attention import BlockMask, flex_attention #KoszarskyB
# -----------------------------------------------------------------------------
# Muon optimizer
@torch.compile
def zeropower_via_newtonschulz5(G, steps=10, eps=1e-7):
"""
Newton-Schulz iteration to compute the zeroth power / orthogonalization of G. We opt to use a
quintic iteration whose coefficients are selected to maximize the slope at zero. For the purpose
of minimizing steps, it turns out to be empirically effective to keep increasing the slope at
zero even beyond the point where the iteration no longer converges all the way to one everywhere
on the interval. This iteration therefore does not produce UV^T but rather something like US'V^T
where S' is diagonal with S_{ii}' ~ Uniform(0.5, 1.5), which turns out not to hurt model
performance at all relative to UV^T, where USV^T = G is the SVD.
"""
assert len(G.shape) == 2
a, b, c = (3.4445, -4.7750, 2.0315)
X = G.bfloat16()
X /= (X.norm() + eps) # ensure top singular value <= 1
if G.size(0) > G.size(1):
X = X.T
for _ in range(steps):
A = X @ X.T
B = b * A + c * A @ A # adapted from suggestion by @jxbz, @leloykun, and @YouJiacheng
X = a * X + B @ X
if G.size(0) > G.size(1):
X = X.T
return X
class Muon(torch.optim.Optimizer):
"""
Muon - MomentUm Orthogonalized by Newton-schulz
Muon internally runs standard SGD-momentum, and then performs an orthogonalization post-
processing step, in which each 2D parameter's update is replaced with the nearest orthogonal
matrix. To efficiently orthogonalize each update, we use a Newton-Schulz iteration, which has
the advantage that it can be stably run in bfloat16 on the GPU.
Some warnings:
- This optimizer assumes that all parameters passed in are 2D.
- It should not be used for the embedding layer, the final fully connected layer, or any {0,1}-D
parameters; those should all be optimized by a standard method (e.g., AdamW).
- To use it with 4D convolutional filters, it works well to just flatten their last 3 dimensions.
- We believe it is unlikely to work well for training with small batch size.
- We believe it may not work well for finetuning pretrained models, but we haven't tested this.
- We have not yet tried this optimizer for training scenarios larger than NanoGPT (124M).
Arguments:
lr: The learning rate used by the internal SGD.
momentum: The momentum used by the internal SGD.
nesterov: Whether to use Nesterov-style momentum in the internal SGD. (recommended)
ns_steps: The number of Newton-Schulz iteration steps to use.
"""
def __init__(self, params, lr=0.02, momentum=0.95, nesterov=True, ns_steps=5):
self.world_size = int(os.environ['WORLD_SIZE'])
self.rank = int(os.environ['RANK'])
defaults = dict(lr=lr, momentum=momentum, nesterov=nesterov, ns_steps=ns_steps)
params = list(params)
assert all(isinstance(p, torch.Tensor) for p in params)
sizes = {p.numel() for p in params}
param_groups = [
{
'params': [p for p in params if p.numel() == size],
'update_buffer': [
torch.empty(size, device='cuda', dtype=torch.bfloat16)
for _ in range(self.world_size)
],
}
for size in sizes
]
super().__init__(param_groups, defaults)
def step(self):
for group in self.param_groups:
lr = group['lr']
momentum = group['momentum']
nesterov = group['nesterov']
ns_steps = group['ns_steps']
update_buffers = group['update_buffer']
# generate weight updates in distributed fashion
params = group['params']
assert len(params) % self.world_size == 0
handle = None
params_world = None
def update_prev():
if params_world is None:
return
assert handle is not None
handle.wait()
for p_world, g_world in zip(params_world, update_buffers):
p_world.data.add_(
g_world.view_as(p_world),
alpha=-lr * max(1, p_world.size(0) / p_world.size(1)) ** 0.5,
)
for base_i in range(len(params))[::self.world_size]:
p = params[base_i + self.rank]
g = p.grad
assert g is not None
state = self.state[p]
if 'momentum_buffer' not in state:
state['momentum_buffer'] = torch.zeros_like(g)
buf = state['momentum_buffer']
buf.lerp_(g, 1 - momentum)
g = g.lerp_(buf, momentum) if nesterov else buf
g = zeropower_via_newtonschulz5(g, steps=ns_steps).flatten()
update_prev()
handle = dist.all_gather(update_buffers, g, async_op=True)
params_world = params[base_i : base_i + self.world_size]
update_prev()
# -----------------------------------------------------------------------------
# PyTorch nn.Module definitions for the GPT-2 model
def norm(x):
return F.rms_norm(x, (x.size(-1),))
class CastedLinear(nn.Linear):
def __init__(self, in_features, out_features):
super().__init__(in_features, out_features, bias=False)
def forward(self, x):
return F.linear(x, self.weight.to(x.dtype))
class Rotary(torch.nn.Module):
def __init__(self, dim, base=10000):
super().__init__()
self.register_buffer('inv_freq', (1 / base) ** (torch.arange(0, dim, 2) / dim))
self.seq_len_cached = None
self.cos_cached = None
self.sin_cached = None
def forward(self, x):
seq_len = x.shape[1]
if seq_len != self.seq_len_cached:
t = torch.arange(seq_len, device=x.device)
freqs = torch.outer(t, self.inv_freq)
self.seq_len_cached = seq_len
self.cos_cached = freqs.cos()
self.sin_cached = freqs.sin()
cos, sin = self.cos_cached[None, :, None, :], self.sin_cached[None, :, None, :]
# apply_rotary_emb(x, cos, sin)
x1, x2 = x.chunk(2, dim=3)
y1 = x1 * cos + x2 * sin
y2 = x1 * (-sin) + x2 * cos
return torch.cat((y1, y2), 3).type_as(x)
class CausalSelfAttention(nn.Module):
def __init__(self, dim, num_heads):
super().__init__()
assert dim % num_heads == 0
self.num_heads = num_heads
self.c_q = CastedLinear(dim, dim)
self.c_k = CastedLinear(dim, dim)
self.c_v = CastedLinear(dim, dim)
self.lambdas = nn.Parameter(torch.tensor([0.5, 0.5]))
self.rotary = Rotary(dim // num_heads) # dim // num_heads = head_dim
self.c_proj = CastedLinear(dim, dim)
self.c_proj.weight.data.zero_() # zero init suggested by @Grad62304977
def forward(self, x, vi, block_mask):
B, T = x.size(0), x.size(1) # batch size, sequence length
assert B == 1, "Must use batch size = 1 for FlexAttention"
q = self.c_q(x).view(B, T, self.num_heads, -1)
k = self.c_k(x).view(B, T, self.num_heads, -1)
v = self.c_v(x).view(B, T, self.num_heads, -1)
v = self.lambdas[0] * v + self.lambdas[1] * vi.view_as(v) # @KoszarskyB & @Grad62304977
q, k = norm(q), norm(k) # QK norm @Grad62304977
q, k = self.rotary(q), self.rotary(k)
y = flex_attention(q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2), block_mask=block_mask, enable_gqa=True)
y = y.transpose(1, 2).contiguous().view_as(x) # re-assemble all head outputs side by side
y = self.c_proj(y)
return y
class MLP(nn.Module):
def __init__(self, dim):
super().__init__()
self.c_fc = CastedLinear(dim, 4 * dim)
self.c_proj = CastedLinear(4 * dim, dim)
self.c_proj.weight.data.zero_() # zero init suggested by @Grad62304977
def forward(self, x):
x = self.c_fc(x)
x = F.relu(x).square() # https://arxiv.org/abs/2109.08668v2; ~1-2% better than GELU; suggested by @SKYLINEZ007 and @Grad62304977
x = self.c_proj(x)
return x
class Block(nn.Module):
def __init__(self, config):
super().__init__()
self.attn = CausalSelfAttention(config.model_dim, config.num_heads)
self.mlp = MLP(config.model_dim)
self.lambdas = nn.Parameter(torch.tensor([1., 0.]))
def forward(self, x, vi, x0, block_mask):
x = self.lambdas[0] * x + self.lambdas[1] * x0
x = x + self.attn(norm(x), vi, block_mask)
x = x + self.mlp(norm(x))
return x
class ValueEmbedding(nn.Module):
def __init__(self, config: "GPTConfig"):
super().__init__()
self.__setattr__
self.embed = nn.ModuleList([
nn.Embedding(config.vocab_size, config.model_dim)
for _ in range(6)
])
def forward(self, inputs) -> "list[torch.Tensor]":
ve = [emb(inputs) for emb in self.embed]
ve += reversed(ve)
return ve
# -----------------------------------------------------------------------------
# The main GPT-2 model
@dataclass
class GPTConfig:
vocab_size : int = 50304
num_layers : int = 12
num_heads : int = 6 # head dim 128 suggested by @Grad62304977
model_dim : int = 768
class GPT(nn.Module):
def __init__(self, config: GPTConfig):
super().__init__()
self.num_layers = config.num_layers
# U-net design by @brendanh0gan
self.num_encoder_layers = config.num_layers // 2 # Half of the layers for encoder
self.num_decoder_layers = config.num_layers - self.num_encoder_layers # Remaining for decoder
# Add learnable skip connection weights for decoder layers
self.skip_weights = nn.Parameter(torch.ones(self.num_decoder_layers))
self.embed = nn.Embedding(config.vocab_size, config.model_dim)
self.blocks = nn.ModuleList([Block(config) for _ in range(config.num_layers)])
# token value embeddings by @KoszarskyB - inspired by @Grad62304977's value residual learning
# U-net structure on token value embeddings by @leloykun
self.value_embeds = ValueEmbedding(config)
self.lm_head = CastedLinear(config.model_dim, config.vocab_size)
self.lm_head.weight.data.zero_() # @Grad62304977
def forward(
self,
inputs: torch.Tensor,
targets: torch.Tensor,
sliding_window_num_blocks: torch.Tensor,
):
BLOCK_SIZE = 128
assert inputs.ndim == 1
docs = (inputs == 50256).cumsum(0)
docs_low = docs.view(-1, BLOCK_SIZE)[:, 0].contiguous()
docs_high = docs.view(-1, BLOCK_SIZE)[:, -1].contiguous()
def document_causal(b, h, q_idx, kv_idx):
causal_mask = q_idx >= kv_idx
document_mask = docs[q_idx] == docs[kv_idx]
return causal_mask & document_mask
def dense_to_ordered(dense_mask: torch.Tensor):
num_blocks = dense_mask.sum(dim=-1, dtype=torch.int32)
indices = dense_mask.argsort(dim=-1, descending=True, stable=True).to(torch.int32)
return num_blocks[None, None].contiguous(), indices[None, None].contiguous()
def create_doc_swc_block_mask(sliding_window_num_blocks: torch.Tensor):
kv_idx = block_idx = torch.arange(512, dtype=torch.int32, device="cuda")
q_idx = block_idx[:, None]
causal_bm = q_idx >= kv_idx
causal_full_bm = q_idx > kv_idx
window_bm = q_idx - kv_idx < sliding_window_num_blocks
window_full_bm = window_bm
# document_bm = (docs_low[q_idx] <= docs_high[kv_idx]) & (docs_low[kv_idx] <= docs_high[q_idx])
document_bm = (docs_low[:, None] <= docs_high) & (docs_low <= docs_high[:, None])
document_full_bm = (docs_low[:, None] == docs_high) & (docs_low == docs_high[:, None])
nonzero_bm = causal_bm & window_bm & document_bm
full_bm = causal_full_bm & window_full_bm & document_full_bm
kv_num_blocks, kv_indices = dense_to_ordered(nonzero_bm ^ full_bm)
full_kv_num_blocks, full_kv_indices = dense_to_ordered(full_bm)
return BlockMask.from_kv_blocks(
kv_num_blocks,
kv_indices,
full_kv_num_blocks,
full_kv_indices,
BLOCK_SIZE=BLOCK_SIZE,
mask_mod=document_causal,
)
block_mask = create_doc_swc_block_mask(sliding_window_num_blocks)
# forward the GPT model itself
x = self.embed(inputs[None]) # token embeddings of shape (b, t, model_dim)
x = norm(x) # @Grad62304977
x0 = x
ve = self.value_embeds(inputs)
ve_enc, ve_dec = ve[:self.num_encoder_layers], ve[self.num_encoder_layers:]
# Store outputs for U-Net skip connections
skip_connections = []
# Encoder pass - process only the first half of the blocks
for i in range(self.num_encoder_layers):
x = self.blocks[i](x, ve_enc[i], x0, block_mask)
skip_connections.append(x)
# Decoder pass - process the remaining blocks with weighted skip connections
for i in range(self.num_decoder_layers):
x = x + self.skip_weights[i] * skip_connections.pop()
# U-net structure on token value embeddings by @leloykun
x = self.blocks[self.num_encoder_layers + i](x, ve_dec[i], x0, block_mask)
x = norm(x)
logits = self.lm_head(x)
logits = 30 * torch.tanh(logits / 30) # @Grad62304977
logits = logits.float()
loss = F.cross_entropy(logits.view(-1, logits.size(-1)), targets.view(-1))
return loss
# -----------------------------------------------------------------------------
# Our own simple Distributed Data Loader
def _peek_data_shard(file: Path):
# only reads the header, returns header data
# header is 256 int32
header = torch.from_file(f"{file}", False, 256, dtype=torch.int32)
assert header[0] == 20240520, "magic number mismatch in the data .bin file"
assert header[1] == 1, "unsupported version"
return int(header[2]) # number of tokens (claimed)
def _load_data_shard(path: Path, num_tokens):
with path.open("rb", buffering=0) as f:
tokens = torch.empty(num_tokens, dtype=torch.uint16, pin_memory=True)
f.seek(256 * 4)
nbytes = f.readinto(tokens.numpy())
assert nbytes == 2 * num_tokens, "number of tokens read does not match header?"
return tokens
class DistributedDataLoader:
def __init__(self, filename_pattern, seq_len, process_rank, num_processes):
self.process_rank = process_rank
self.num_processes = num_processes
self.seq_len = seq_len
# glob files that match the pattern
self.files = sorted(Path.cwd().glob(filename_pattern))
assert len(self.files) > 0, f"did not find any files that match the pattern {filename_pattern}"
# load and validate all data shards, count number of tokens in total
self.files_num_tokens = [_peek_data_shard(file) for file in self.files]
assert min(self.files_num_tokens) >= num_processes * seq_len + 1
self.total_num_tokens = sum(self.files_num_tokens)
self.reset()
def reset(self):
self.current_shard = -1
self.advance()
def advance(self): # advance to next data shard
self.current_shard = (self.current_shard + 1) % len(self.files)
self.current_position = self.process_rank * self.seq_len
self.tokens = _load_data_shard(self.files[self.current_shard], self.files_num_tokens[self.current_shard])
def next_batch(self):
batch_size = self.seq_len * self.num_processes
buf = self.tokens[self.current_position:self.current_position+self.seq_len+1]
# host side async is sufficient;
# no performance improvement was observed when introducing a separate stream.
inputs = buf[:-1].to(device="cuda", dtype=torch.int32, non_blocking=True) # inputs
targets = buf[1:].to(device="cuda", dtype=torch.int64, non_blocking=True) # targets
# advance current position and load next shard if necessary
self.current_position += batch_size
if self.current_position + batch_size + 1 >= len(self.tokens):
self.advance()
return inputs, targets
# -----------------------------------------------------------------------------
# int main
@dataclass
class Hyperparameters:
# data hyperparams
input_bin : str = 'data/fineweb10B/fineweb_train_*.bin' # input .bin to train on
input_val_bin : str = 'data/fineweb10B/fineweb_val_*.bin' # input .bin to eval validation loss on
# optimization hyperparams
batch_size : int = 8 # batch size, in sequences, across all devices
sequence_length : int = 64*1024 # sequence length, in tokens
num_iterations : int = 1480 # number of iterations to run
warmup_iters : int = 0
cooldown_iters : int = 600 # number of iterations of linear warmup/cooldown for triangular or trapezoidal schedule
weight_decay : float = 0
# evaluation and logging hyperparams
val_loss_every : int = 125 # every how many steps to evaluate val loss? 0 for only at the end
val_tokens : int = 10485760 # how many tokens of validation data? it's important to keep this fixed for consistent comparisons
save_every : int = 0 # every how many steps to save the checkpoint? 0 for only at the end
args = Hyperparameters()
# set up DDP (distributed data parallel). torchrun sets this env variable
ddp_rank = int(os.environ['RANK'])
ddp_local_rank = int(os.environ['LOCAL_RANK'])
ddp_world_size = int(os.environ['WORLD_SIZE'])
assert torch.cuda.is_available()
device = torch.device(f"cuda:{ddp_local_rank}")
torch.cuda.set_device(device)
print(f"using device: {device}")
dist.init_process_group(backend='nccl', device_id=device)
dist.barrier()
master_process = (ddp_rank == 0) # this process will do logging, checkpointing etc.
# begin logging
logfile = None
if master_process:
run_id = uuid.uuid4()
logdir = Path("logs") / f"{run_id}"
logdir.mkdir(exist_ok=True)
logfile = Path("logs") / f"{run_id}.txt"
print(logfile.stem)
# create the log file
with logfile.open("w") as f:
# begin the log by printing this file (the Python code)
print(code, file=f)
print("=" * 100, file=f)
def print0(s, logonly=False):
if master_process:
with logfile.open("a") as f:
if not logonly:
print(s)
print(s, file=f)
# log information about the hardware/software environment this is running on
# and print the full `nvidia-smi` to file
print0(f"Running python {sys.version}")
print0(f"Running pytorch {torch.version.__version__} compiled for CUDA {torch.version.cuda}\nnvidia-smi:")
import subprocess
result = subprocess.run(['nvidia-smi'], stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True)
print0(f'{result.stdout}', logonly=True)
print0('='*100, logonly=True)
# calculate the number of steps to take in the val loop.
assert args.val_tokens % (args.sequence_length * ddp_world_size) == 0
val_steps = args.val_tokens // (args.sequence_length * ddp_world_size)
# calculate the steps of gradient accumulation required to attain the desired global batch size.
assert args.batch_size % (ddp_world_size) == 0
train_accumulation_steps = args.batch_size // ddp_world_size
# load tokens
train_loader = DistributedDataLoader(args.input_bin, args.sequence_length, ddp_rank, ddp_world_size)
val_loader = DistributedDataLoader(args.input_val_bin, args.sequence_length, ddp_rank, ddp_world_size)
print0(f"Training DataLoader: total number of tokens: {train_loader.total_num_tokens} across {len(train_loader.files)} files")
print0(f"Validation DataLoader: total number of tokens: {val_loader.total_num_tokens} across {len(val_loader.files)} files")
print0('='*100, logonly=True)
inputs_train, targets_train = train_loader.next_batch()
# there are only 50257 unique GPT-2 tokens; we extend to nearest multiple of 128 for efficiency. suggested to me by @Grad62304977.
# this originates from Karpathy's experiments.
num_vocab = 50304
model = GPT(GPTConfig(vocab_size=num_vocab, num_layers=12, num_heads=6, model_dim=768))
model = model.cuda().bfloat16()
for m in model.modules():
if isinstance(m, CastedLinear):
m.float()
config.coordinate_descent_tuning = True # suggested by @Chillee
model = torch.compile(model)
# here we wrap model into DDP container
model = DDP(model, device_ids=[ddp_local_rank], broadcast_buffers=False, gradient_as_bucket_view=True)
raw_model = model.module # always contains the "raw" unwrapped model
# init the optimizer(s)
embed_params = [*raw_model.embed.parameters(), *raw_model.value_embeds.parameters()]
optimizer1 = torch.optim.Adam(embed_params, lr=0.6, betas=(0.8, 0.95), fused=True)
optimizer2 = torch.optim.Adam([raw_model.lm_head.weight], lr=0.008, betas=(0.8, 0.95), fused=True)
params = list(raw_model.blocks.parameters())
matrix_params = [p for p in params if p.ndim == 2]
scalar_params = [p for p in params if p.ndim < 2] + [raw_model.skip_weights]
optimizer3 = Muon(matrix_params, lr=0.05, momentum=0.95)
optimizer4 = torch.optim.Adam(scalar_params, lr=0.04, betas=(0.8, 0.95), fused=True)
optimizers = [optimizer1, optimizer2, optimizer3, optimizer4]
# learning rate decay scheduler (linear warmup and cooldown)
def get_lr(it):
assert it <= args.num_iterations
# 1) linear warmup for warmup_iters steps
if it < args.warmup_iters:
return (it+1) / args.warmup_iters
# 2) constant lr for a while
elif it < args.num_iterations - args.cooldown_iters:
return 1.0
# 3) linear cooldown
else:
decay_ratio = (args.num_iterations - it) / args.cooldown_iters
return decay_ratio
schedulers = [torch.optim.lr_scheduler.LambdaLR(opt, get_lr) for opt in optimizers]
sliding_window_num_blocks = torch.tensor(1, dtype=torch.int32, device="cuda")
sw_num_blocks_prev = 1
# Start training loop
training_time_ms = 0
# start the clock
torch.cuda.synchronize()
t0 = time.perf_counter()
# begin training
for step in range(args.num_iterations + 1):
last_step = (step == args.num_iterations)
# This effectively ignores timing first 10 steps, which are slower for weird reasons.
# Alternately, and slightly more correctly in terms of benchmarking, we could do 10
# steps with dummy data first, and then re-initialize the model and reset the loader.
if step == 10:
training_time_ms = 0
t0 = time.perf_counter()
timed_steps = float('nan') if step <= 11 else (step - 10) + 1 # <= 11 to avoid bug in val
# Linearly increase the sliding window size over training in chunks of 64 from 64 -> 1792. By @fernbear.bsky.social
frac_done = step / args.num_iterations # training progress
sw_num_blocks = int(((1 - frac_done) * 64 + frac_done * 1792 + 64) // 128)
if sw_num_blocks != sw_num_blocks_prev:
sliding_window_num_blocks.copy_(sw_num_blocks, non_blocking=True)
sw_num_blocks_prev = sw_num_blocks
# once in a while evaluate the validation dataset
if (last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0)):
# stop the clock
torch.cuda.synchronize()
training_time_ms += 1000 * (time.perf_counter() - t0)
# run validation batches
model.eval()
val_loader.reset()
val_loss = 0.0
for _ in range(val_steps):
with torch.no_grad():
inputs_val, targets_val = val_loader.next_batch()
val_loss += model(inputs_val, targets_val, sliding_window_num_blocks)
dist.all_reduce(val_loss, op=dist.ReduceOp.AVG)
val_loss /= val_steps
# log val loss to console and to logfile
print0(f'step:{step}/{args.num_iterations} val_loss:{val_loss:.4f} train_time:{training_time_ms:.0f}ms step_avg:{training_time_ms/(timed_steps-1):.2f}ms')
# start the clock again
torch.cuda.synchronize()
t0 = time.perf_counter()
if master_process and (last_step or (args.save_every > 0 and step % args.save_every == 0)):
# stop the clock
torch.cuda.synchronize()
training_time_ms += 1000 * (time.perf_counter() - t0)
# save the state of the training process
log = dict(step=step, code=code, model=raw_model.state_dict(), optimizers=[opt.state_dict() for opt in optimizers])
torch.save(log, 'logs/%s/state_step%06d.pt' % (run_id, step))
# start the clock again
torch.cuda.synchronize()
t0 = time.perf_counter()
# bit confusing: we want to make sure to eval on 0th iteration
# but also after the very last iteration. so we loop for step <= num_iterations
# instead of just < num_iterations (one extra due to <=), only to do
# the validation/sampling one last time, and then we break right here as we're done.
if last_step:
break
# --------------- TRAINING SECTION BEGIN -----------------
model.train()
for i in range(1, train_accumulation_steps + 1):
with contextlib.ExitStack() as stack:
if i < train_accumulation_steps: # there's no need to sync gradients every accumulation step
stack.enter_context(model.no_sync())
if step >= 5:
stack.enter_context(torch.compiler.set_stance(skip_guard_eval_unsafe=True))
model(inputs_train, targets_train, sliding_window_num_blocks).backward()
inputs_train, targets_train = train_loader.next_batch()
if train_accumulation_steps != 1:
for p in model.parameters():
p.grad /= train_accumulation_steps
# momentum warmup for Muon
frac = min(step/300, 1)
for group in optimizer3.param_groups:
group['momentum'] = (1 - frac) * 0.85 + frac * 0.95
# step the optimizers and schedulers
for opt, sched in zip(optimizers, schedulers):
opt.step()
sched.step()
# null the gradients
model.zero_grad(set_to_none=True)
# --------------- TRAINING SECTION END -------------------
# everything that follows now is just diagnostics, prints, logging, etc.
approx_time = training_time_ms + 1000 * (time.perf_counter() - t0)
print0(f"step:{step+1}/{args.num_iterations} train_time:{approx_time:.0f}ms step_avg:{approx_time/timed_steps:.2f}ms")
print0(f"peak memory consumption: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB")
# -------------------------------------------------------------------------
# clean up nice
dist.destroy_process_group()
====================================================================================================
Running python 3.10.12 (main, Sep 11 2024, 15:47:36) [GCC 11.4.0]
Running pytorch 2.6.0.dev20241203+cu124 compiled for CUDA 12.4
nvidia-smi:
Wed Dec 11 09:29:38 2024
+---------------------------------------------------------------------------------------+
| NVIDIA-SMI 535.183.06 Driver Version: 535.183.06 CUDA Version: 12.2 |
|-----------------------------------------+----------------------+----------------------+
| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|=========================================+======================+======================|
| 0 NVIDIA H100 80GB HBM3 On | 00000000:19:00.0 Off | 0 |
| N/A 38C P0 125W / 700W | 7084MiB / 81559MiB | 0% Default |
| | | Disabled |
+-----------------------------------------+----------------------+----------------------+
| 1 NVIDIA H100 80GB HBM3 On | 00000000:3B:00.0 Off | 0 |
| N/A 30C P0 116W / 700W | 3451MiB / 81559MiB | 0% Default |
| | | Disabled |
+-----------------------------------------+----------------------+----------------------+
| 2 NVIDIA H100 80GB HBM3 On | 00000000:4C:00.0 Off | 0 |
| N/A 29C P0 111W / 700W | 3451MiB / 81559MiB | 0% Default |
| | | Disabled |
+-----------------------------------------+----------------------+----------------------+
| 3 NVIDIA H100 80GB HBM3 On | 00000000:5D:00.0 Off | 0 |
| N/A 36C P0 114W / 700W | 3451MiB / 81559MiB | 0% Default |
| | | Disabled |
+-----------------------------------------+----------------------+----------------------+
| 4 NVIDIA H100 80GB HBM3 On | 00000000:9B:00.0 Off | 0 |
| N/A 38C P0 119W / 700W | 3451MiB / 81559MiB | 0% Default |
| | | Disabled |
+-----------------------------------------+----------------------+----------------------+
| 5 NVIDIA H100 80GB HBM3 On | 00000000:BB:00.0 Off | 0 |
| N/A 30C P0 118W / 700W | 3451MiB / 81559MiB | 0% Default |
| | | Disabled |
+-----------------------------------------+----------------------+----------------------+
| 6 NVIDIA H100 80GB HBM3 On | 00000000:CB:00.0 Off | 0 |
| N/A 36C P0 119W / 700W | 3451MiB / 81559MiB | 0% Default |
| | | Disabled |
+-----------------------------------------+----------------------+----------------------+
| 7 NVIDIA H100 80GB HBM3 On | 00000000:DB:00.0 Off | 0 |
| N/A 30C P0 118W / 700W | 3211MiB / 81559MiB | 0% Default |
| | | Disabled |
+-----------------------------------------+----------------------+----------------------+
+---------------------------------------------------------------------------------------+
| Processes: |
| GPU GI CI PID Type Process name GPU Memory |
| ID ID Usage |
|=======================================================================================|
+---------------------------------------------------------------------------------------+
====================================================================================================
Training DataLoader: total number of tokens: 1000000000 across 10 files
Validation DataLoader: total number of tokens: 100000000 across 1 files
====================================================================================================
step:0/1480 val_loss:10.8258 train_time:0ms step_avg:nanms
step:1/1480 train_time:28925ms step_avg:nanms
step:2/1480 train_time:29102ms step_avg:nanms
step:3/1480 train_time:29226ms step_avg:nanms
step:4/1480 train_time:29365ms step_avg:nanms
step:5/1480 train_time:29507ms step_avg:nanms
step:6/1480 train_time:29650ms step_avg:nanms
step:7/1480 train_time:29791ms step_avg:nanms
step:8/1480 train_time:29934ms step_avg:nanms
step:9/1480 train_time:30076ms step_avg:nanms
step:10/1480 train_time:30224ms step_avg:nanms
step:11/1480 train_time:136ms step_avg:nanms
step:12/1480 train_time:276ms step_avg:nanms
step:13/1480 train_time:420ms step_avg:139.91ms
step:14/1480 train_time:564ms step_avg:140.97ms
step:15/1480 train_time:707ms step_avg:141.35ms
step:16/1480 train_time:849ms step_avg:141.55ms
step:17/1480 train_time:990ms step_avg:141.49ms
step:18/1480 train_time:1134ms step_avg:141.73ms
step:19/1480 train_time:1277ms step_avg:141.85ms
step:20/1480 train_time:1421ms step_avg:142.07ms
step:21/1480 train_time:1563ms step_avg:142.08ms
step:22/1480 train_time:1706ms step_avg:142.18ms
step:23/1480 train_time:1849ms step_avg:142.24ms
step:24/1480 train_time:1990ms step_avg:142.11ms
step:25/1480 train_time:2132ms step_avg:142.10ms
step:26/1480 train_time:2274ms step_avg:142.15ms
step:27/1480 train_time:2418ms step_avg:142.26ms
step:28/1480 train_time:2563ms step_avg:142.38ms
step:29/1480 train_time:2706ms step_avg:142.43ms
step:30/1480 train_time:3217ms step_avg:160.85ms
step:31/1480 train_time:3317ms step_avg:157.95ms
step:32/1480 train_time:3461ms step_avg:157.32ms
step:33/1480 train_time:3603ms step_avg:156.66ms
step:34/1480 train_time:3746ms step_avg:156.07ms
step:35/1480 train_time:3886ms step_avg:155.45ms
step:36/1480 train_time:4028ms step_avg:154.93ms
step:37/1480 train_time:4171ms step_avg:154.50ms
step:38/1480 train_time:4316ms step_avg:154.14ms
step:39/1480 train_time:4461ms step_avg:153.84ms
step:40/1480 train_time:4606ms step_avg:153.52ms
step:41/1480 train_time:4749ms step_avg:153.18ms
step:42/1480 train_time:4890ms step_avg:152.82ms
step:43/1480 train_time:5033ms step_avg:152.50ms
step:44/1480 train_time:5177ms step_avg:152.26ms
step:45/1480 train_time:5320ms step_avg:152.00ms
step:46/1480 train_time:5463ms step_avg:151.76ms
step:47/1480 train_time:5607ms step_avg:151.55ms
step:48/1480 train_time:5749ms step_avg:151.28ms
step:49/1480 train_time:5890ms step_avg:151.02ms
step:50/1480 train_time:6032ms step_avg:150.81ms
step:51/1480 train_time:6176ms step_avg:150.62ms
step:52/1480 train_time:6320ms step_avg:150.47ms
step:53/1480 train_time:6464ms step_avg:150.32ms
step:54/1480 train_time:6607ms step_avg:150.15ms
step:55/1480 train_time:6749ms step_avg:149.97ms
step:56/1480 train_time:6891ms step_avg:149.80ms
step:57/1480 train_time:7034ms step_avg:149.66ms
step:58/1480 train_time:7176ms step_avg:149.50ms
step:59/1480 train_time:7320ms step_avg:149.40ms
step:60/1480 train_time:7465ms step_avg:149.29ms
step:61/1480 train_time:7607ms step_avg:149.17ms
step:62/1480 train_time:7749ms step_avg:149.03ms
step:63/1480 train_time:7892ms step_avg:148.90ms
step:64/1480 train_time:8034ms step_avg:148.78ms
step:65/1480 train_time:8179ms step_avg:148.72ms
step:66/1480 train_time:8323ms step_avg:148.63ms
step:67/1480 train_time:8465ms step_avg:148.52ms
step:68/1480 train_time:8608ms step_avg:148.42ms
step:69/1480 train_time:8751ms step_avg:148.31ms
step:70/1480 train_time:8892ms step_avg:148.19ms
step:71/1480 train_time:9033ms step_avg:148.08ms
step:72/1480 train_time:9175ms step_avg:147.99ms
step:73/1480 train_time:9320ms step_avg:147.93ms
step:74/1480 train_time:9463ms step_avg:147.86ms
step:75/1480 train_time:9606ms step_avg:147.79ms
step:76/1480 train_time:9747ms step_avg:147.69ms
step:77/1480 train_time:9889ms step_avg:147.59ms
step:78/1480 train_time:10030ms step_avg:147.51ms
step:79/1480 train_time:10549ms step_avg:152.89ms
step:80/1480 train_time:10650ms step_avg:152.14ms
step:81/1480 train_time:10791ms step_avg:151.98ms
step:82/1480 train_time:10932ms step_avg:151.84ms
step:83/1480 train_time:11075ms step_avg:151.71ms
step:84/1480 train_time:11217ms step_avg:151.59ms
step:85/1480 train_time:11360ms step_avg:151.47ms
step:86/1480 train_time:11504ms step_avg:151.37ms
step:87/1480 train_time:11648ms step_avg:151.28ms
step:88/1480 train_time:11790ms step_avg:151.15ms
step:89/1480 train_time:11933ms step_avg:151.05ms
step:90/1480 train_time:12076ms step_avg:150.95ms
step:91/1480 train_time:12219ms step_avg:150.85ms
step:92/1480 train_time:12362ms step_avg:150.76ms
step:93/1480 train_time:12505ms step_avg:150.66ms
step:94/1480 train_time:12648ms step_avg:150.57ms
step:95/1480 train_time:12790ms step_avg:150.47ms
step:96/1480 train_time:12932ms step_avg:150.37ms
step:97/1480 train_time:13465ms step_avg:154.77ms
step:98/1480 train_time:13566ms step_avg:154.15ms
step:99/1480 train_time:13709ms step_avg:154.03ms
step:100/1480 train_time:13850ms step_avg:153.89ms
step:101/1480 train_time:13996ms step_avg:153.80ms
step:102/1480 train_time:14134ms step_avg:153.63ms
step:103/1480 train_time:14276ms step_avg:153.51ms
step:104/1480 train_time:14420ms step_avg:153.40ms
step:105/1480 train_time:14563ms step_avg:153.29ms
step:106/1480 train_time:14706ms step_avg:153.18ms
step:107/1480 train_time:14848ms step_avg:153.07ms
step:108/1480 train_time:14989ms step_avg:152.95ms
step:109/1480 train_time:15130ms step_avg:152.82ms
step:110/1480 train_time:15272ms step_avg:152.72ms
step:111/1480 train_time:15417ms step_avg:152.64ms
step:112/1480 train_time:15564ms step_avg:152.59ms
step:113/1480 train_time:15710ms step_avg:152.52ms
step:114/1480 train_time:15854ms step_avg:152.44ms
step:115/1480 train_time:16000ms step_avg:152.38ms
step:116/1480 train_time:16147ms step_avg:152.33ms
step:117/1480 train_time:16291ms step_avg:152.25ms
step:118/1480 train_time:16438ms step_avg:152.20ms
step:119/1480 train_time:16583ms step_avg:152.14ms
step:120/1480 train_time:16729ms step_avg:152.08ms
step:121/1480 train_time:16874ms step_avg:152.02ms
step:122/1480 train_time:17021ms step_avg:151.97ms
step:123/1480 train_time:17167ms step_avg:151.92ms
step:124/1480 train_time:17312ms step_avg:151.86ms
step:125/1480 train_time:17458ms step_avg:151.81ms
step:125/1480 val_loss:4.4180 train_time:17523ms step_avg:152.38ms
step:126/1480 train_time:17616ms step_avg:151.86ms
step:127/1480 train_time:17762ms step_avg:151.81ms
step:128/1480 train_time:17909ms step_avg:151.77ms
step:129/1480 train_time:18054ms step_avg:151.71ms
step:130/1480 train_time:18198ms step_avg:151.65ms
step:131/1480 train_time:18345ms step_avg:151.61ms
step:132/1480 train_time:18491ms step_avg:151.56ms
step:133/1480 train_time:18635ms step_avg:151.50ms
step:134/1480 train_time:18783ms step_avg:151.48ms
step:135/1480 train_time:18931ms step_avg:151.45ms
step:136/1480 train_time:19076ms step_avg:151.39ms
step:137/1480 train_time:19221ms step_avg:151.35ms
step:138/1480 train_time:19368ms step_avg:151.31ms
step:139/1480 train_time:19513ms step_avg:151.27ms
step:140/1480 train_time:19658ms step_avg:151.22ms
step:141/1480 train_time:19804ms step_avg:151.17ms
step:142/1480 train_time:19950ms step_avg:151.14ms
step:143/1480 train_time:20095ms step_avg:151.09ms
step:144/1480 train_time:20240ms step_avg:151.04ms
step:145/1480 train_time:20386ms step_avg:151.01ms
step:146/1480 train_time:20531ms step_avg:150.96ms
step:147/1480 train_time:20676ms step_avg:150.92ms
step:148/1480 train_time:20823ms step_avg:150.89ms
step:149/1480 train_time:20969ms step_avg:150.86ms
step:150/1480 train_time:21114ms step_avg:150.81ms
step:151/1480 train_time:21260ms step_avg:150.78ms
step:152/1480 train_time:21407ms step_avg:150.76ms
step:153/1480 train_time:21552ms step_avg:150.72ms
step:154/1480 train_time:21696ms step_avg:150.67ms
step:155/1480 train_time:21843ms step_avg:150.64ms
step:156/1480 train_time:21989ms step_avg:150.61ms
step:157/1480 train_time:22134ms step_avg:150.57ms
step:158/1480 train_time:22279ms step_avg:150.53ms
step:159/1480 train_time:22425ms step_avg:150.50ms
step:160/1480 train_time:22571ms step_avg:150.48ms
step:161/1480 train_time:22716ms step_avg:150.44ms
step:162/1480 train_time:22862ms step_avg:150.41ms
step:163/1480 train_time:23009ms step_avg:150.38ms
step:164/1480 train_time:23154ms step_avg:150.35ms
step:165/1480 train_time:23298ms step_avg:150.31ms
step:166/1480 train_time:23445ms step_avg:150.29ms
step:167/1480 train_time:23591ms step_avg:150.26ms
step:168/1480 train_time:23735ms step_avg:150.22ms
step:169/1480 train_time:23881ms step_avg:150.20ms
step:170/1480 train_time:24026ms step_avg:150.17ms
step:171/1480 train_time:24172ms step_avg:150.14ms
step:172/1480 train_time:24316ms step_avg:150.10ms
step:173/1480 train_time:24462ms step_avg:150.07ms
step:174/1480 train_time:24609ms step_avg:150.06ms
step:175/1480 train_time:24754ms step_avg:150.02ms
step:176/1480 train_time:24899ms step_avg:150.00ms
step:177/1480 train_time:25046ms step_avg:149.97ms
step:178/1480 train_time:25192ms step_avg:149.95ms
step:179/1480 train_time:25337ms step_avg:149.92ms
step:180/1480 train_time:25484ms step_avg:149.91ms
step:181/1480 train_time:25630ms step_avg:149.89ms
step:182/1480 train_time:25777ms step_avg:149.87ms
step:183/1480 train_time:25923ms step_avg:149.85ms
step:184/1480 train_time:26071ms step_avg:149.83ms
step:185/1480 train_time:26215ms step_avg:149.80ms
step:186/1480 train_time:26362ms step_avg:149.78ms
step:187/1480 train_time:26508ms step_avg:149.76ms
step:188/1480 train_time:26654ms step_avg:149.74ms
step:189/1480 train_time:26828ms step_avg:149.87ms
step:190/1480 train_time:26944ms step_avg:149.69ms
step:191/1480 train_time:27090ms step_avg:149.67ms
step:192/1480 train_time:27234ms step_avg:149.64ms
step:193/1480 train_time:27379ms step_avg:149.61ms
step:194/1480 train_time:27525ms step_avg:149.59ms
step:195/1480 train_time:27671ms step_avg:149.57ms
step:196/1480 train_time:27816ms step_avg:149.55ms
step:197/1480 train_time:27960ms step_avg:149.52ms
step:198/1480 train_time:28108ms step_avg:149.51ms
step:199/1480 train_time:28254ms step_avg:149.49ms
step:200/1480 train_time:28398ms step_avg:149.46ms
step:201/1480 train_time:28545ms step_avg:149.45ms
step:202/1480 train_time:28691ms step_avg:149.43ms
step:203/1480 train_time:28835ms step_avg:149.40ms
step:204/1480 train_time:28981ms step_avg:149.39ms
step:205/1480 train_time:29127ms step_avg:149.37ms
step:206/1480 train_time:29273ms step_avg:149.35ms
step:207/1480 train_time:29418ms step_avg:149.33ms
step:208/1480 train_time:29564ms step_avg:149.31ms
step:209/1480 train_time:29711ms step_avg:149.30ms
step:210/1480 train_time:29856ms step_avg:149.28ms
step:211/1480 train_time:30000ms step_avg:149.26ms
step:212/1480 train_time:30148ms step_avg:149.25ms
step:213/1480 train_time:30294ms step_avg:149.23ms
step:214/1480 train_time:30438ms step_avg:149.20ms
step:215/1480 train_time:30584ms step_avg:149.19ms
step:216/1480 train_time:30730ms step_avg:149.17ms
step:217/1480 train_time:30877ms step_avg:149.16ms
step:218/1480 train_time:31023ms step_avg:149.15ms
step:219/1480 train_time:31169ms step_avg:149.13ms
step:220/1480 train_time:31314ms step_avg:149.11ms
step:221/1480 train_time:31874ms step_avg:151.06ms
step:222/1480 train_time:31981ms step_avg:150.85ms
step:223/1480 train_time:32130ms step_avg:150.84ms
step:224/1480 train_time:32278ms step_avg:150.83ms
step:225/1480 train_time:32426ms step_avg:150.82ms
step:226/1480 train_time:32574ms step_avg:150.81ms
step:227/1480 train_time:32722ms step_avg:150.79ms
step:228/1480 train_time:32872ms step_avg:150.79ms
step:229/1480 train_time:33021ms step_avg:150.78ms
step:230/1480 train_time:33169ms step_avg:150.77ms
step:231/1480 train_time:33317ms step_avg:150.76ms
step:232/1480 train_time:33467ms step_avg:150.75ms
step:233/1480 train_time:33615ms step_avg:150.74ms
step:234/1480 train_time:33764ms step_avg:150.73ms
step:235/1480 train_time:33913ms step_avg:150.73ms
step:236/1480 train_time:34062ms step_avg:150.72ms
step:237/1480 train_time:34212ms step_avg:150.71ms
step:238/1480 train_time:34360ms step_avg:150.70ms
step:239/1480 train_time:34509ms step_avg:150.69ms
step:240/1480 train_time:34657ms step_avg:150.68ms
step:241/1480 train_time:34805ms step_avg:150.67ms
step:242/1480 train_time:34954ms step_avg:150.67ms
step:243/1480 train_time:35102ms step_avg:150.65ms
step:244/1480 train_time:35251ms step_avg:150.65ms
step:245/1480 train_time:35399ms step_avg:150.63ms
step:246/1480 train_time:35549ms step_avg:150.63ms
step:247/1480 train_time:35697ms step_avg:150.62ms
step:248/1480 train_time:35846ms step_avg:150.61ms
step:249/1480 train_time:35995ms step_avg:150.61ms
step:250/1480 train_time:36142ms step_avg:150.59ms
step:250/1480 val_loss:3.9974 train_time:36209ms step_avg:150.87ms
step:251/1480 train_time:36301ms step_avg:150.63ms
step:252/1480 train_time:36450ms step_avg:150.62ms
step:253/1480 train_time:36599ms step_avg:150.61ms
step:254/1480 train_time:36747ms step_avg:150.60ms
step:255/1480 train_time:36896ms step_avg:150.60ms
step:256/1480 train_time:37044ms step_avg:150.59ms
step:257/1480 train_time:37193ms step_avg:150.58ms
step:258/1480 train_time:37341ms step_avg:150.57ms
step:259/1480 train_time:37493ms step_avg:150.57ms
step:260/1480 train_time:37641ms step_avg:150.56ms
step:261/1480 train_time:37789ms step_avg:150.56ms
step:262/1480 train_time:37938ms step_avg:150.55ms
step:263/1480 train_time:38086ms step_avg:150.54ms
step:264/1480 train_time:38235ms step_avg:150.53ms
step:265/1480 train_time:38384ms step_avg:150.53ms
step:266/1480 train_time:38533ms step_avg:150.52ms
step:267/1480 train_time:38681ms step_avg:150.51ms
step:268/1480 train_time:38829ms step_avg:150.50ms
step:269/1480 train_time:38977ms step_avg:150.49ms
step:270/1480 train_time:39124ms step_avg:150.48ms
step:271/1480 train_time:39274ms step_avg:150.47ms
step:272/1480 train_time:39421ms step_avg:150.46ms
step:273/1480 train_time:39571ms step_avg:150.46ms
step:274/1480 train_time:39719ms step_avg:150.45ms
step:275/1480 train_time:39870ms step_avg:150.45ms
step:276/1480 train_time:40018ms step_avg:150.44ms
step:277/1480 train_time:40166ms step_avg:150.43ms
step:278/1480 train_time:40315ms step_avg:150.43ms
step:279/1480 train_time:40463ms step_avg:150.42ms
step:280/1480 train_time:40612ms step_avg:150.42ms
step:281/1480 train_time:40760ms step_avg:150.41ms
step:282/1480 train_time:40909ms step_avg:150.40ms
step:283/1480 train_time:41058ms step_avg:150.39ms
step:284/1480 train_time:41206ms step_avg:150.39ms
step:285/1480 train_time:41354ms step_avg:150.38ms
step:286/1480 train_time:41503ms step_avg:150.37ms
step:287/1480 train_time:41652ms step_avg:150.37ms
step:288/1480 train_time:41800ms step_avg:150.36ms
step:289/1480 train_time:41949ms step_avg:150.36ms
step:290/1480 train_time:42098ms step_avg:150.35ms
step:291/1480 train_time:42247ms step_avg:150.34ms
step:292/1480 train_time:42395ms step_avg:150.34ms
step:293/1480 train_time:42543ms step_avg:150.33ms
step:294/1480 train_time:42693ms step_avg:150.33ms
step:295/1480 train_time:42841ms step_avg:150.32ms
step:296/1480 train_time:42991ms step_avg:150.32ms
step:297/1480 train_time:43139ms step_avg:150.31ms
step:298/1480 train_time:43288ms step_avg:150.31ms
step:299/1480 train_time:43437ms step_avg:150.30ms
step:300/1480 train_time:43585ms step_avg:150.29ms
step:301/1480 train_time:43734ms step_avg:150.29ms
step:302/1480 train_time:43881ms step_avg:150.28ms
step:303/1480 train_time:44030ms step_avg:150.27ms
step:304/1480 train_time:44179ms step_avg:150.27ms
step:305/1480 train_time:44326ms step_avg:150.26ms
step:306/1480 train_time:44475ms step_avg:150.25ms
step:307/1480 train_time:44622ms step_avg:150.24ms
step:308/1480 train_time:44771ms step_avg:150.24ms
step:309/1480 train_time:44920ms step_avg:150.23ms
step:310/1480 train_time:45068ms step_avg:150.23ms
step:311/1480 train_time:45217ms step_avg:150.22ms
step:312/1480 train_time:45364ms step_avg:150.21ms
step:313/1480 train_time:45513ms step_avg:150.21ms
step:314/1480 train_time:45661ms step_avg:150.20ms
step:315/1480 train_time:45809ms step_avg:150.19ms
step:316/1480 train_time:45956ms step_avg:150.18ms
step:317/1480 train_time:46104ms step_avg:150.18ms
step:318/1480 train_time:46253ms step_avg:150.17ms
step:319/1480 train_time:46401ms step_avg:150.16ms
step:320/1480 train_time:46550ms step_avg:150.16ms
step:321/1480 train_time:46699ms step_avg:150.16ms
step:322/1480 train_time:46847ms step_avg:150.15ms
step:323/1480 train_time:46996ms step_avg:150.15ms
step:324/1480 train_time:47143ms step_avg:150.14ms
step:325/1480 train_time:47294ms step_avg:150.14ms
step:326/1480 train_time:47441ms step_avg:150.13ms
step:327/1480 train_time:47591ms step_avg:150.13ms
step:328/1480 train_time:47739ms step_avg:150.12ms
step:329/1480 train_time:47888ms step_avg:150.12ms